Amazon Web Services released a new Step Functions pattern this week that assigns AI agents a role in airline rebooking while ensuring that reservation modifications and payment transactions remain under code's direct control. The pattern orchestrates Amazon Bedrock AgentCore agents to propose alternative itineraries and write compensation messages following flight disruptions, but deterministic workflow steps must validate every suggestion before any reservation changes or payments execute. According to AWS, the guiding principle is straightforward: "agents propose, and deterministic code validates."
The pattern uses AWS Step Functions to wrap each agent's non-deterministic behavior with deterministic validation steps, preventing any agent from taking direct actions such as writing reservations or issuing payments. Instead, agent proposals only take effect after passing deterministic validation, with Step Functions maintaining an execution history for audit purposes. This contrasts with multi-agent collaboration frameworks, where a supervisor agent orchestrates sub-agent runs and tool calls—AWS's approach moves those decisions out of the agent layer entirely and into the Step Functions workflow. On the same day, AWS published a case study of Abnormal AI, a behavioral AI security platform that uses Amazon Bedrock AgentCore Code Interpreter for real-time inline email threat detection. The Code Interpreter provides a fully managed, serverless runtime where agents can execute code dynamically, pairing what AWS describes as the semantic reasoning of a large language model with the calculation, data processing, and verification that come from running code.
AWS writes that the separation between agent reasoning and deterministic validation could help developers deploy AI agents more safely in high-stakes scenarios. The company positions Code Interpreter not simply as a coding tool but as "fundamental infrastructure that agents use to reason computationally." For tasks that don't map cleanly to semantic reasoning—converting data into structured reports, for instance, or counting—a compute scratchpad lets agents work through problems computationally and verify answers rather than relying on reasoning alone. The report notes that airline rebooking is well-suited to agentic workflows because agents can offload the tedious work of finding route alternatives, comparing constraints, and coordinating next steps, freeing operations teams from manually clicking through hundreds of passenger rebookings after a flight cancellation.
AWS argues that as agents handle more consequential tasks—spanning accounts payable, procurement, and the monthly close—semantic reasoning alone can't guarantee that agent proposals are safe enough to execute. A layer of deterministic validation may at least provide a verifiable checkpoint between proposal and action, and the same pattern could extend to other high-stakes financial and regulatory workflows where deterministic code should stand between agent suggestions and execution. AWS isn't alone in pursuing this separation: last month, Perplexity shipped Portable Computer, the local-first version of its Computer agent running on an Nvidia DGX Spark workstation that also uses deterministic software to control model actions, separating probabilistic reasoning from execution decisions. Organizations deploying agents in environments where errors carry financial or compliance risk will need to weigh the operational flexibility that agents promise against the imperative to maintain human or code-level control over irreversible decisions. The pattern AWS describes offers one architectural answer, though the broader question of where to draw the line between proposal and execution remains open territory as agent capabilities expand.